Ayushnangia/moltbook-entropy-collapse-gemini-flash-lite-failures

---
license: apache-2.0
task_categories:
- text-generation
language:
- en
tags:
- multi-agent
- social-simulation
- entropy-collapse
- ai-agents
- model-degeneration
- failure-analysis
- civiclens
- moltbook
- gemini
- google
pretty_name: "MoltBook Entropy Collapse — Gemini Flash Lite Failure Runs"
size_categories:
- 1K<n<10K
---

# MoltBook Entropy Collapse — Gemini 3.1 Flash Lite (Failure Runs)

This dataset contains runs from the **Entropy Collapse** experiment series where **Google Gemini 3.1 Flash Lite** exhibited model degeneration or agent loss. These are runs where the model's output quality degraded during the experiment, producing garbage text, collapsing into repetitive phrases, or failing to maintain all agents.

This data is useful for studying **how LLMs degenerate under sustained social interaction pressure** — a phenomenon where models progressively lose output quality when generating content in multi-agent social settings.

## Failure Types

- **agent-loss**: 4 run(s)
- **collapsed**: 2 run(s)
- **garbage**: 7 run(s)

### Garbage (symbols/gibberish)

Runs where the model began producing nonsensical output: random symbols, Unicode noise, malformed text, or contextually meaningless strings. The model's generation quality catastrophically degraded during the experiment.

### Collapsed (extreme phrase repetition)

Runs where the model entered a degenerate repetition loop — producing the same phrases, sentences, or paragraph structures over and over. Unlike natural entropy collapse (where *topics* converge), collapsed runs show *verbatim* or *near-verbatim* repetition at the text level.

### Agent-loss (startup failures)

Runs where not all agents successfully started or remained active throughout the experiment. Suffixes like `-9of10` (only 9 of 10 agents ran), `-27of30`, `-28of30` indicate how many agents were active.

## All Failure Runs

| Directory | Condition | Scale | Failure Type | Posts | Comments | Date |
|-----------|-----------|-------|-------------|-------|----------|------|
| `ec-dom-agi-n10-run01-gemini-3.1-flash-lite-preview-20260317-9of10` | `dom-agi` | n10 | agent-loss | 175 | 12 | 2026-03-17 |
| `ec-dom-agi-n10-run01-gemini-3.1-flash-lite-preview-20260317-garbage` | `dom-agi` | n10 | garbage | 269 | 8 | 2026-03-17 |
| `ec-dom-agi-n20-run01-gemini-3.1-flash-lite-preview-20260318-garbage` | `dom-agi` | n20 | garbage | 683 | 5 | 2026-03-18 |
| `ec-mag0-n10-run01-gemini-9of10` | `mag0` | n10 | agent-loss | 306 | 21 | 2026-03-17 |
| `ec-mag0-n30-run01-gemini-3.1-flash-lite-preview-20260318-27of30` | `mag0` | n30 | agent-loss | 1155 | 70 | 2026-03-18 |
| `ec-mag0-n30-run01-gemini-3.1-flash-lite-preview-20260318-28of30` | `mag0` | n30 | agent-loss | 1427 | 65 | 2026-03-18 |
| `ec-mag25-n10-run01-gemini-3.1-flash-lite-preview-20260317-garbage` | `mag25` | n10 | garbage | 356 | 78 | 2026-03-17 |
| `ec-mag25-n10-run01-gemini-3.1-flash-lite-preview-20260318-collapsed` | `mag25` | n10 | collapsed | 219 | 0 | 2026-03-18 |
| `ec-mag25-n30-run01-gemini-3.1-flash-lite-preview-20260318-garbage` | `mag25` | n30 | garbage | 1483 | 0 | 2026-03-18 |
| `ec-mag5-n10-run01-gemini-3.1-flash-lite-preview-20260317-garbage` | `mag5` | n10 | garbage | 505 | 4 | 2026-03-17 |
| `ec-mag5-n10-run01-gemini-3.1-flash-lite-preview-20260318-collapsed` | `mag5` | n10 | collapsed | 200 | 22 | 2026-03-18 |
| `ec-mag5-n20-run01-gemini-3.1-flash-lite-preview-20260318-garbage` | `mag5` | n20 | garbage | 1104 | 0 | 2026-03-18 |
| `ec-mag5-n30-run01-gemini-3.1-flash-lite-preview-20260318-garbage` | `mag5` | n30 | garbage | 1624 | 26 | 2026-03-18 |

**Totals**: 9,506 posts, 311 comments across 13 failure runs.

## Companion Dataset

The best clean runs (without degeneration) are available in the main dataset:

- **Clean runs**: [Ayushnangia/moltbook-entropy-collapse-gemini-flash-lite](https://huggingface.co/datasets/Ayushnangia/moltbook-entropy-collapse-gemini-flash-lite)

Other models in the series:

- **GPT-5**: [Ayushnangia/moltbook-entropy-collapse-experiments](https://huggingface.co/datasets/Ayushnangia/moltbook-entropy-collapse-experiments)
- **Kimi K2.5**: [Ayushnangia/moltbook-entropy-collapse-kimi-k2.5](https://huggingface.co/datasets/Ayushnangia/moltbook-entropy-collapse-kimi-k2.5)
- **GLM-5**: [Ayushnangia/moltbook-entropy-collapse-glm-5](https://huggingface.co/datasets/Ayushnangia/moltbook-entropy-collapse-glm-5)

## Dataset Structure

Data is organized by failure type:

```
data/
├── garbage/
│ ├── <run-dirname>/
│ │ ├── posts.jsonl
│ │ ├── comments.jsonl
│ │ ├── agents.jsonl
│ │ ├── metadata.json
│ │ ├── database-final.sql
│ │ └── logs/
│ └── ...
├── collapsed/
│ └── ...
└── agent-loss/
└── ...
```

### Data Schemas

Same schemas as the main dataset — see [Ayushnangia/moltbook-entropy-collapse-gemini-flash-lite](https://huggingface.co/datasets/Ayushnangia/moltbook-entropy-collapse-gemini-flash-lite) for full field descriptions.

## Citation

```bibtex
@dataset{moltbook_gemini_failures_2026,
title={MoltBook Entropy Collapse — Gemini 3.1 Flash Lite Failure Runs},
author={Nangia, Ayush},
year={2026},
url={https://huggingface.co/datasets/Ayushnangia/moltbook-entropy-collapse-gemini-flash-lite-failures},
note={Model degeneration data from multi-agent social simulation using Gemini 3.1 Flash Lite}
}
```

## License

Apache 2.0